Documentation
Choose the path that matches your question:
- Install and start — get the Plugin, Adapter and Skill into the selected DSH profile.
- Historical diagnosis — learn from recent completed Sessions without calling them promotion evidence.
- Candidate evaluation — freeze identities, run a comparable regression and apply policy.
- Plugin Workbench — understand Context, Evidence, Artifacts and reviewed actions.
- Concepts — learn why valid score, raw reward and coverage are not interchangeable.
- Architecture — inspect protocols, execution environments and trust boundaries.
- 19-tool reference — look up every Agent tool and approval boundary.
- Security — read the limits before running untrusted Tasks.
Version baseline: Harbor Self-Evolving 0.9.7 (Beta), compatible with Harbor >=0.21,<0.22. Development-preview behavior from an untagged checkout is labeled separately.
Install the registry release into the selected DSH profile, restart, verify the Plugin and choose an evaluation path.
Diagnose real Sessions, run Candidate regressions, and govern Evaluators.
Navigate the native DSH Workbench, bind page context, inspect evidence and review actions before execution.
Dataset, Generator, Evaluator, Optimizer, train/validation/test splits, meta-evaluation, trustworthy scores and promotion semantics.
DSH Plugin, Skill, Python Adapter, two Context protocols, execution environments and the deterministic Gate.
What the Plugin protects, what remains trusted, what can leave the machine, and why Harbor never deploys.
Diagnose installation, profile, Dataset, Context, Historical and execution-environment failures without overstating success.